he social media owner Meta stated on Friday that they are launching some new artificial intelligence models, including a “Self-Taught Evaluator”
system that will provide a way to involve less human input in the AI development process.
This launch comes months after the first announcement in an August paper, which stated that this AI system uses the same “chain of thought” approach that OpenAI uses in its recent o1 models to make honest judgments in the responses.
The OpenAI technique works by breaking down major and complex problems into smaller and logical steps that are specially created to improve the authenticity of the responses on different issues in science, coding, or math.
For the social media metaverse version of the Meta AI tool, researchers used data and information that were fully AI-generated to train the evaluator AI system to eliminate human input in all stages of the development process. This ability to use artificial intelligence to examine AI systems seems very innovative because this might offer a new way of building Meta AI tools that are capable of learning from their mistakes.
Subscribe to our newsletter
Soon, these self-improving AI systems will eventually eliminate the need to use expensive and inefficient processes such as the one that is used today, the Reinforcement Learning from Human Feedback. This process requires human input from reviewers with specialized expertise to analyze data and information accurately along with examining answers to determine if some complex math and writing interrogations are indeed correct.
"We hope, as AI becomes more and more super-human, that it will get better and better at checking its work so that it will actually be better than the average human," said Jason Weston, one of the researchers., one of the researchers, John Weston said about this new Meta AI tool.
"The idea of being self-taught and able to self-evaluate is basically crucial to the idea of getting to this sort of super-human level of AI," he also said in his statement.
This Friday, the owner of the social media platform Facebook also launched other AI systems that are based on Meta’s image identification Segment Anything model. This tool is capable of speeding up the LLM response times along with the datasets to be used in the discovery process of some inorganic materials.
Also, other companies such as Google and Anthropic have published recently some research on Reinforcement learning from AI feedback, also known as the RLAIF. However, these companies are not planning to release near in the future artificial intelligence models open to the public.
